First Assessment Report - IPCC

248 Dctec tion of the Gieenhouse Effec t in the Ohsei \ations 8
current scientific focus in the detection issue is thereloic on
multivariate or fingerprint analyses The fingerprint method
is essentially a form of model validation, wheie the
perturbation experiment that is being used to test the
models is the currently uncontrolled emission of
greenhouse gases into the atmosphere The method is
discussed further in Section 8 3 First, however, we
consider some of the more general issues of a detection
strategy
8.2 Detection Strategies
8.2.1 Choosing Detection Variables
There are many possible climate elements or sets of
elements that we could study to try to detect an enhanced
greenhouse effect In choosing the ones to study, the
following issues must be considered
the strength of the predicted signal and the ease with
which it may be distinguished from the noise,
uncertainties in both the predicted signal and the noise,
and
the availability and quality of suitable observed data
8 2 11 Signal to noise tatios
The signal-to-noise ratio provides a convenient criterion for
ranking different possible detection variables The stronger
the predicted signal relative to the noise, the better the
variable will be for detection purposes, all other things
being equal For multivariate signals, those for which the
pattern of natural variability is distinctly different from the
pattern of the predicted signal will automatically have a
high signal-to noise ratio
Signal to-noise ratios have been calculated tor a number
of individual climate elements from the results of lxCCb
and 2xCC>2 equilibrium experiments using atmospheric
GCMs coupled to mixed-layei oceans (Bainett and
Schlesinger 1987, Santer et al , 1990 Schlesinger ct al
1990) The highest values were obtained lor free
troposphere temperatures, near-surface tempeiatures
(including sea-surface temperatures), and lower to middle
troposphenc water vapour content (especially in tropical
regions) Lowest values were iound for mean sea level
pressure and precipitation While these results may be
model dependent, they do provide a useful preliminary
indicator of the relative values of different elements in the
detection context
Variables with distinctly difleient signal and noise
patterns may be difficult to find (Bamett and Schlesinger,
1987) There are reasons to expect parallels between the
signal and the low-frequency noise patterns at least at the
zonal and seasonal levels, simply because such char
actenstics anse through feedback mechanisms that are
common to both greenhouse forcing and natural variability
82 I 2 Signal umeitauities
Clearly a vanable toi which the signal is highly uncertain
cannot be a good candidate as a detection variable
Predicted signals depend on the models used to produce
them Model-to-model differences (Section 5) point
strongly to laige signal uncertainties Some insights into
these uncertainties may also be gained from studies ol
model results in attempting to simulate the present-day
climate (see Section 4) A poor representation of the
present climate would indicate greater uncertainty in the
predicted signal (e g , Mitchell et al , 1987) Such
uncertainties tend to be largest at the regional scale because
the processes that act on these scales are not accurately
represented or paiametenzed in the models Even if a
particular model is able to simulate the present-day climate
well, it will still be difficult to estimate how well it can
define an enhanced greenhouse signal Nevertheless,
validations of simulations ol the present global climate
should form at least one of the bases for the selection of
detection variables
A source of unccitainty hcie is the difference between
the results of equilibrium and transient experiments (see
Section 6) Studies using coupled ocean-atmosphere GCMs
and time-varying CO2 loicing have shown reduced
warming in the aieas of deep water formation (1 e , the
North Atlantic basin and around Antarctica) compared with
equilibrium results (Bryan ct al , 1988, Washington and
Mcehl, 1989, Stouffer et al , 1989) These experiments
suggest that the regional patterns of temperature change
may be more complex than those predicted by equilibrium
simulations The results of equilibrium experiments must
therefore be considered as only a guide to possible signal
structuie
The most reliable signals aie likely to be those related to
the largest spatial scales Small-scale details may be
eliminated by spatial averaging, or, more generally by
using filters that pass only the larger scale (low wave
number) components (Note that some relatively smallscale
features may be appropriate for detection purposes, if
model confidence is high ) An additional benefit of spatial
averaging or filtenng is that it results in data compression
(1 e , reducing the dimensionality of the detection variable),
which facilitates statistical testing Data compression may
also be achieved by using linear combinations of variables
(e g , Bell, 1982, 1986, Kaioly, 1987, 1989)
8 2 13 Noise unteitamties
Since the expected man-made climatic changes occur on
decadal and longer time-scales, it is largely the lowfrequency
characteristics of natural variability that are
important in defining the noise Estimating the magnitude
of low frequency variability presents a major problem
because of the shoitness and incompleteness of most